Guides · 2026-10-11
China AI Labeling: Explicit vs Implicit Labels
While the EU was finalizing the AI Act, China moved first on labeling: its AI-generated content identification measures took effect on 1 September 2025, requiring both visible and embedded labels on synthetic content. If you publish anything AI-generated that reaches Chinese audiences — or distribute through Chinese platforms — here's what the rules demand.
Two kinds of labels: explicit and implicit
The measures define two complementary labeling layers. Explicit labels are what users see: visible text, badges or voice notices identifying content as AI-generated, placed where users can't miss them. Implicit labels are what systems read: metadata, digital watermarks or fingerprints embedded in the file that identify the content as synthetic without altering the user experience.
Sound familiar? It maps almost exactly onto the EU AI Act's visible-disclosure plus machine-readable-marking split. The global direction is unmistakable: every major jurisdiction is converging on the same two-layer model.
Who has to label
The duties fall primarily on service providers — the platforms and AI services operating in China — which must add labels to AI-generated content on their services and provide labeling capabilities to users. But users and distributors carry duties too: anyone distributing AI-generated content must not strip, alter or conceal the labels, and must add explicit labels when publishing content they know (or should know) is AI-generated.
The anti-stripping rule is the sharp edge. Removing an AI label before reposting isn't a gray area under these measures — it's a violation. If your workflow involves downloading, editing and re-uploading content, every handoff is a point where labels can die, and the liability follows the content.
What "explicit" means in practice
For text: a visible notice at the beginning, end or another prominent position. For images and video: a prominent visible mark — on-image text or badges that survive normal viewing. For audio: a voice notice at the start, end or appropriate intervals. The standard is prominence: a label a user would have to work to miss, not one they'd have to work to find.
What "implicit" means in practice
Implicit labeling covers the machine layer: metadata fields, digital watermarks, content fingerprints — anything that lets platforms and regulators identify synthetic content programmatically. Providers are expected to use robust techniques that survive ordinary processing. In practice this means the same toolkit the EU expects: C2PA-style manifests, watermarking, generator metadata.
Enforcement and penalties
Enforcement runs through China's cyberspace and telecom regulators, with the standard administrative toolkit: orders to correct, warnings, fines, and — for serious cases — service suspensions. Platforms face particular pressure: app stores and distribution channels are expected to police labeling on the content they carry, which pushes compliance duties down the supply chain to every publisher.
Practical steps for international publishers
If your content reaches China — through TikTok/Douyin, Xiaohongshu, WeChat Channels, or any Chinese distribution partner — build labeling into the production workflow, not the publishing step. Add explicit labels during editing (burned-in text or badges survive every platform's recompression), preserve whatever implicit metadata your tools emit, and brief every editor and freelancer on the anti-stripping rule: re-exporting without labels is the violation, even when it was an accident. For teams publishing across jurisdictions, the efficient move is to label once to the strictest standard you face — content labeled for China's explicit-label rules will generally satisfy the EU and Google Ads visible-disclosure duties too.
The global compliance picture
Step back and the pattern is clear. The EU (Article 50, applicable August 2026), China (labeling measures, effective September 2025) and the major ad platforms (Google, Meta disclosure rules) all require the same two layers: visible disclosure for humans, machine-readable marking for systems, applied at publication, never stripped in distribution.
That convergence is good news for publishers: one disciplined workflow covers all three regimes. Label visibly at first exposure, embed machine-readable provenance, keep per-asset records, and verify labels survived publishing. The specifics differ — wording, placement rules, penalty levels — but the operating discipline is identical. DisclosureGuard checks assets against versioned rule packs for the EU, US/Google Ads and China in a single scan, so one workflow produces evidence for every jurisdiction you publish in.
Informational explainer, not legal advice. Check a specific asset against the current rules with a free DisclosureGuard scan.